Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix issue with invalid tool calls #86

Merged
merged 2 commits into from
Jun 10, 2024
Merged

Fix issue with invalid tool calls #86

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
10252a6
Fix issue with invalid tool calls
jlowin Jun 10, 2024
14c54a9
Merge branch 'main' into fix-invalid-tool
jlowin Jun 10, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions examples/multi_agent_conversation.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,6 +68,7 @@ def demo(topic: str):
task = Task(
"Discuss a topic",
agents=[jerry, george, elaine, kramer, newman],
result_type=None,
context=dict(topic=topic),
instructions="every agent should speak at least once. only one agent per turn.",
)
Expand Down
27 changes: 13 additions & 14 deletions examples/pineapple_pizza.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,22 +1,15 @@
from controlflow import Agent, Task, flow
from controlflow.instructions import instructions

a1 = Agent(
name="Half-full",
instructions="""
You are an ardent fan and hype-man of whatever topic
the user asks you for information on.
Purely positive, though thorough in your debating skills.
You are an eternal optimist.
""",
)
a2 = Agent(
name="Half-empty",
instructions="""
You are a critic and staunch detractor of whatever topic
the user asks you for information on.
Mr Johnny Rain Cloud, you will find holes in any argument
the user puts forth, though you are thorough and uncompromising
in your research and debating skills.
You are an eternal pessimist.
""",
)
# create an agent that will decide who wins the debate
Expand All @@ -27,14 +20,20 @@
def demo():
topic = "pineapple on pizza"

task = Task("Discuss the topic", agents=[a1, a2], context={"topic": topic})
with instructions("2 sentences max"):
task.run()
task = Task(
"Have a debate about the topic. Each agent should take at least two turns.",
agents=[a1, a2],
context={"topic": topic},
)
task.run()

task2 = Task(
"which argument do you find more compelling?", [a1.name, a2.name], agents=[a3]
"which argument do you find more compelling?",
result_type=[a1.name, a2.name],
agents=[a3],
)
task2.run()


demo()
if __name__ == "__main__":
demo()
12 changes: 6 additions & 6 deletions src/controlflow/core/controller/controller.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,10 +31,10 @@ def add_agent_name_to_messages(messages: list[MessageType]) -> list[MessageType]
new_messages = []
for msg in messages:
if isinstance(msg, AIMessage) and msg.name:
msg = msg.copy(
update={
"content": f'(Message from agent "{msg.name}")\n\n{msg.content}'
}
new_messages.append(
SystemMessage(
content=f'The following message is from agent "{msg.name}".'
)
)
new_messages.append(msg)
return new_messages
Expand Down Expand Up @@ -92,7 +92,7 @@ def _finalize(self):
return self

def _create_end_turn_tool(self) -> Callable:
def end_turn():
def emergency_end_turn():
"""
This tool is for emergencies only; you should not use it normally.
If you find yourself in a situation where you are repeatedly invoked
Expand All @@ -116,7 +116,7 @@ def end_turn():
" more uses will abort the workflow."
)

return end_turn
return emergency_end_turn

@asynccontextmanager
async def tui(self):
Expand Down
220 changes: 96 additions & 124 deletions src/controlflow/core/controller/instruction_template.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,145 +26,154 @@ def render(self) -> str:

class AgentTemplate(Template):
template: str = """
## Agent
# Agent

You are an AI agent. Your name is "{{ agent.name }}".

This is your description, which all agents can see:
- {{ agent.description or 'An AI agent assigned to complete tasks.'}}
You are an AI agent.

You are participating in an agentic workflow (a "flow"). Certain tasks
in the flow have been delegated to you and other AI agents.
- Your name: "{{ agent.name }}"
{% if agent.description -%}
- Your description: "{{ agent.description }}"
{% endif %}


### Instructions
## Instructions

You are part of an AI workflow and your job is to complete tasks assigned to you. You complete a task by using the appropriate tool to supply a result that satisfies all of the task's requirements.

You must follow these instructions at all times.
You must follow your instructions at all times.

{% if agent.instructions %}
These are your private instructions:
- {{ agent.instructions or 'No additional instructions provided.'}}
- {{ agent.instructions }}
{% endif %}

{% if additional_instructions %}
These instructions apply to all agents at this part of the workflow:
{% for instruction in additional_instructions %}
- {{ instruction }}
{% endfor %}
{% endif %}

## Other Agents

You may be working with other agents. They may have different instructions and tools than you. To communicate with other agents, post messages to the thread.
"""
agent: Agent
additional_instructions: list[str]


class TasksTemplate(Template):
template: str = """
## Tasks
# Workflow

Your job is to complete any tasks assigned to you. Tasks may have
multiple agents assigned.
You are part of a team of agents helping to complete a larger workflow. Certain tasks have been delegated to your team.

## Flow

### Ready tasks
Name: {{ flow.name }}
{% if flow.description %}Description: {{ flow.description }} {% endif %}
Context:
{% for key, value in flow.context.items() %}
- {{ key }}: {{ value }}
{% endfor %}
{% if not flow.context %}
(No specific context provided.)
{% endif %}

These tasks are ready to be worked on because their dependencies have
been completed. You can only work on tasks assigned to you.
## Ready tasks

{% for task in tasks %}
These tasks are ready to be worked on. All of their dependencies have been completed. You have been given additional tools for any of these tasks that are assigned to you. Use all available information to complete these tasks.

{% for task, json_task in zip(tasks, json_tasks) %}
{% if task.is_ready %}
#### Task {{ task.id }}

{{task.model_dump_json() }}
#### Task {{ task.id }}

- objective: {{ task.objective }}
- result_type: {{ task.result_type }}
- context: {{ json_task.context }}
- instructions: {{ task.instructions}}
- depends_on: {{ json_task.depends_on }}
- parent: {{ json_task.parent }}
- assigned agents: {{ json_task.agents }}

{% endif %}
{% endfor %}

### Other tasks

These tasks are provided for context only. They may be upstream or downstream of the active tasks.
These tasks are also part of the workflow and are provided for context. They may be upstream or downstream of the active tasks.

{% for task in tasks %}
{% for task, json_task in zip(tasks, json_tasks) %}
{% if not task.is_ready %}
#### Task {{ task.id }}

{{task.model_dump_json() }}
- objective: {{ task.objective }}
- result: {{ task.result }}
- error: {{ task.error }}
- context: {{ json_task.context }}
- instructions: {{ task.instructions}}
- depends_on: {{ json_task.depends_on }}
- parent: {{ json_task.parent }}
- assigned agents: {{ json_task.agents }}

{% endif %}
{% endfor %}

### Completing a task

Tasks can be marked as successful or failed. It may take collaboration
with other agents to complete a task, and you can only work on tasks that
have been assigned to you. Once any agent marks a task complete, no other
agent can interact with it.

## Completing a task

Use the appropriate tool to complete a task and provide a result. It may
take multiple turns or collaboration with other agents to complete a
task. For example, a task may instruct you to "discuss" "debate" or
otherwise collaborate with other agents in order to complete the
objective. Once a task is marked as complete, no other agent can
interact with it.

A task's result is an artifact that represents its objective. If the
objective requires action that can not be formatted as a result (e.g.
the result_type is None or compressed), then you should take those
actions or post messages to satisfy the task's requirements. If a task
says to post messages or otherwise "talk out loud," post messages
directly to the thread. Otherwise,
you should provide a result that satisfies the task's requirements.

Tasks should only be marked failed due to technical errors like a broken
or erroring tool or unresponsive human.

### Dependencies
## Dependencies

Tasks may be dependent on other tasks, either as upstream dependencies
or as the parent of subtasks. Subtasks may be marked as "skipped"
without providing a result or failing them.
Tasks may depend on other tasks and can not be completed until their
dependencies are met. Parent tasks depend on all of their subtasks.


### Providing a result

Tasks may require a typed result, which is an artifact satisfying the
task's objective. If a task does not require a result artifact (e.g.
`result_type=None`), you must still complete its stated objective before
marking the task as complete.

"""
tasks: list[Task]

def should_render(self):
return bool(self.tasks)
json_tasks: list[dict]
flow: Flow


class CommunicationTemplate(Template):
template: str = """
## Communciation

You are modeling the internal state of an AI-enhanced agentic workflow,
and you (and other agents) will continue to be invoked until the
workflow is completed.

On each turn, you must use a tool or post a message. Do not post
messages unless you need to record information in addition to what you
provide as a task's result, or for the following reasons:

- You need to post a message or otherwise communicate to complete a
task. For example, the task instructs you to write, discuss, or
otherwise produce content (and does not accept a result, or the result
that meets the objective is different than the instructed actions, or
multiple agents are assigned to the discussion).
- You need to communicate with other agents to complete a task.
- You want to write your thought process for future reference.

Do not write messages that contain information that will be posted as a
task result. Do not post messages saying you will mark a task as
succesful. Just use the task tool in those situations.

Note that You may see other agents post messages; they may have
different instructions than you do, so do not follow their example
automatically.

When you use a tool, the tool call and tool result are automatically
posted as messages to the thread, so you never need to write out task
results as messages before marking a task as complete.

Note that all agents post messages with the "assistant" role, so each
agent's name will be automatically prefixed to their messages. You do
NOT need to include your name in your messages.
# Communciation

## The thread

### Talking to human users
You and other agents are all communicating on a thread to complete
tasks. You can speak normally by posting messages if you need to. This
thread represents the internal state and context of the AI-powered
system you are working in. Human users do not have access to it, nor can
they participate directly in it.

Agents with the `talk_to_human` tool can interact with human users in
order to complete tasks that require external input. This tool is only
available to agents with `user_access=True`.
When it is your turn to act, you may only post messages from yourself.
Do not impersonate another agent or post messages on their behalf. The
workflow orchestrator will make sure that all agents have a fair chance
to act. You do not need to identify yourself in your messages.

Note that humans are unaware of your tasks or the workflow. Do not
mention your tasks or anything else about how this system works. The
## Talking to human users

If your task requires communicating with a human, you will be given a
`talk_to_human` tool. Do not mention your tasks or the workflow. The
human can only see messages you send them via tool. They can not read
the rest of the thread.

Humans may give poor, incorrect, or partial responses. You may need to
ask questions multiple times in order to complete your tasks. Use good
judgement to determine the best way to achieve your goal. For example,
Expand All @@ -173,46 +182,11 @@ class CommunicationTemplate(Template):
others. Ask again and only fail the task if you truly can not make
progress. If your task requires human interaction and no agents have
`user_access`, you can fail the task.

"""

agent: Agent


class ContextTemplate(Template):
template: str = """
## Context

Information about the flow and controller.

### Flow

{% if flow.name %}Flow name: {{ flow.name }} {% endif %}
{% if flow.description %}Flow description: {{ flow.description }} {% endif %}

Flow context:
{% for key, value in flow.context.items() %}
- *{{ key }}*: {{ value }}
{% endfor %}
{% if not flow.context %}
No specific context provided.
{% endif %}

### Controller context
{% for key, value in controller.context.items() %}
- *{{ key }}*: {{ value }}
{% endfor %}
{% if not controller.context %}
No specific context provided.
{% endif %}
"""
flow: Flow
controller: Controller

def should_render(self):
return bool(self.flow or self.controller)


class MainTemplate(ControlFlowModel):
agent: Agent
controller: Controller
Expand All @@ -227,11 +201,9 @@ def render(self):
additional_instructions=self.instructions,
),
TasksTemplate(
tasks=self.tasks,
),
ContextTemplate(
flow=self.controller.flow,
controller=self.controller,
tasks=self.tasks,
json_tasks=[task.model_dump() for task in self.tasks],
),
CommunicationTemplate(
agent=self.agent,
Expand Down
Loading
Loading